Signal attenuation enables scalable decentralized multi-agent reinforcement learning over networks